Machine-Learning Classification of Soil Bulk Density in Salt Marsh Environments.

Area of Science:

• Environmental science
• Remote sensing
• Soil science

Background:

• Remotely sensed data (visible, near-infrared, shortwave infrared) are cost-effective for soil property characterization.
• Estimating salt marsh soil bulk density using remote sensing at the vegetation rooting zone is an underexplored area.

Purpose of the Study:

• To assess machine learning algorithms (Random Forest, XGBoost, SVM) for modeling salt marsh soil bulk density.
• To evaluate the utility of Landsat-7 ETM+ multispectral data for this estimation.

Main Methods:

• Applied Random Forest (RF), Extreme Gradient Boosting Machines (XGBoost), and Support Vector Machines (SVM) to Landsat-7 ETM+ data.
• Utilized visible (blue) and near-infrared (NIR) spectral bands as key predictors.
• Classified soil bulk density into low (0.032-0.752 g/cm³) and high (0.752-1.893 g/cm³) categories.

Main Results:

• XGBoost and RF identified blue and NIR bands as most important for bulk density prediction.
• XGBoost achieved the highest classification accuracy (88%), closely followed by RF (87%) and SVM (86%).
• XGBoost accurately classified 96% of low bulk density and 60% of high bulk density samples.

Conclusions:

• Remote-sensing-based machine learning models are effective for mapping wetland soil bulk density.
• This approach provides valuable data for ecologists and engineers in site selection for wetland restoration.
• The study highlights the potential of Landsat-7 data for understanding soil properties critical for vegetation re-establishment.